Design And Implementation Of Voice Wake-up System For Voiceprint Recognition Based On Deep Learning

With the rapid development of artificial intelligence technology,a variety of intelligent terminal devices have been launched in the market.The development of intelligent voice makes voice control intelligent terminal equipment become the main technology research direction of major intelligent terminal equipment enterprises,Voice wake-up is the entrance of intelligent terminal equipment and user interaction.How to efficiently and accurately respond to the user's voice signal input containing keywords has become the most important goal of this technology.However,pure voice wake-up does not guarantee our personal information security.As the entrance of human-computer interaction,we need to add identification voiceprint recognition technology.Voiceprint recognition is that users input voice to intelligent terminal devices and match voiceprint features through algorithms.Voice wake-up refers to that when users say specific voice commands,the device switches from sleep state to working state,Give the specified response.Voice wake-up task can be regarded as a kind of keyword detection task with small resources,which takes up less computing resources and CPU space.Therefore,the framework of keyword detection system is quite different from that of speech recognition system.Voiceprint recognition task,through the literal understanding of this is a kind of biometric recognition technology,also known as target speaker recognition model,there are two types of target speaker recognition and target speaker recognition.The purpose of this paper is to design and implement a voice wake-up system for voiceprint recognition based on deep learning.By combining the technology innovation of voiceprint recognition and voice wake-up,and comparing the deep neural network architecture,the main work is as follows:(1)Referring to the end-to-end model architecture,separate training is adopted for voiceprint recognition module and voice wake-up module,so that the accuracy and wake-up rate of the two models are integrated to improve the accuracy and interactive effect of the whole system.(2)In the voiceprint recognition module,the combination of the end-to-end matching model LSTM and DNN is used,and the combination of linear and nonlinear activation functions is used to make the model have a more accurate matching process and lower parameter calculation.(3)In the voice wake-up module,the end-to-end Wave Net model is used as the network architecture of voice wake-up.The advantages of the transfer learning and expansion convolution and residual network are that the initial parameters of speech recognition and the lower parameter calculation of the expanded convolution and the residual information of the residual network are used to make the input prediction better timing.The whole network can ensure that we have a good understanding of the model Low power consumption and small resource requirements.Finally,through the comparative experiment on the model network selection,we test the different contrast effects of voiceprint recognition and voice wake-up models and network selection.The system trains a large number of Chinese and English Parallel Corpus through two modules of voiceprint recognition and voice wake-up,so that the accuracy of the model can be guaranteed,and the accuracy of voice recognition to voice wake-up can be accurately and accurately Through the integration of the two models,the original intention of the system is realized.